xref: /openbmc/linux/drivers/net/wireguard/send.c (revision dc6a81c3)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2015-2019 Jason A. Donenfeld <Jason@zx2c4.com>. All Rights Reserved.
4  */
5 
6 #include "queueing.h"
7 #include "timers.h"
8 #include "device.h"
9 #include "peer.h"
10 #include "socket.h"
11 #include "messages.h"
12 #include "cookie.h"
13 
14 #include <linux/uio.h>
15 #include <linux/inetdevice.h>
16 #include <linux/socket.h>
17 #include <net/ip_tunnels.h>
18 #include <net/udp.h>
19 #include <net/sock.h>
20 
21 static void wg_packet_send_handshake_initiation(struct wg_peer *peer)
22 {
23 	struct message_handshake_initiation packet;
24 
25 	if (!wg_birthdate_has_expired(atomic64_read(&peer->last_sent_handshake),
26 				      REKEY_TIMEOUT))
27 		return; /* This function is rate limited. */
28 
29 	atomic64_set(&peer->last_sent_handshake, ktime_get_coarse_boottime_ns());
30 	net_dbg_ratelimited("%s: Sending handshake initiation to peer %llu (%pISpfsc)\n",
31 			    peer->device->dev->name, peer->internal_id,
32 			    &peer->endpoint.addr);
33 
34 	if (wg_noise_handshake_create_initiation(&packet, &peer->handshake)) {
35 		wg_cookie_add_mac_to_packet(&packet, sizeof(packet), peer);
36 		wg_timers_any_authenticated_packet_traversal(peer);
37 		wg_timers_any_authenticated_packet_sent(peer);
38 		atomic64_set(&peer->last_sent_handshake,
39 			     ktime_get_coarse_boottime_ns());
40 		wg_socket_send_buffer_to_peer(peer, &packet, sizeof(packet),
41 					      HANDSHAKE_DSCP);
42 		wg_timers_handshake_initiated(peer);
43 	}
44 }
45 
46 void wg_packet_handshake_send_worker(struct work_struct *work)
47 {
48 	struct wg_peer *peer = container_of(work, struct wg_peer,
49 					    transmit_handshake_work);
50 
51 	wg_packet_send_handshake_initiation(peer);
52 	wg_peer_put(peer);
53 }
54 
55 void wg_packet_send_queued_handshake_initiation(struct wg_peer *peer,
56 						bool is_retry)
57 {
58 	if (!is_retry)
59 		peer->timer_handshake_attempts = 0;
60 
61 	rcu_read_lock_bh();
62 	/* We check last_sent_handshake here in addition to the actual function
63 	 * we're queueing up, so that we don't queue things if not strictly
64 	 * necessary:
65 	 */
66 	if (!wg_birthdate_has_expired(atomic64_read(&peer->last_sent_handshake),
67 				      REKEY_TIMEOUT) ||
68 			unlikely(READ_ONCE(peer->is_dead)))
69 		goto out;
70 
71 	wg_peer_get(peer);
72 	/* Queues up calling packet_send_queued_handshakes(peer), where we do a
73 	 * peer_put(peer) after:
74 	 */
75 	if (!queue_work(peer->device->handshake_send_wq,
76 			&peer->transmit_handshake_work))
77 		/* If the work was already queued, we want to drop the
78 		 * extra reference:
79 		 */
80 		wg_peer_put(peer);
81 out:
82 	rcu_read_unlock_bh();
83 }
84 
85 void wg_packet_send_handshake_response(struct wg_peer *peer)
86 {
87 	struct message_handshake_response packet;
88 
89 	atomic64_set(&peer->last_sent_handshake, ktime_get_coarse_boottime_ns());
90 	net_dbg_ratelimited("%s: Sending handshake response to peer %llu (%pISpfsc)\n",
91 			    peer->device->dev->name, peer->internal_id,
92 			    &peer->endpoint.addr);
93 
94 	if (wg_noise_handshake_create_response(&packet, &peer->handshake)) {
95 		wg_cookie_add_mac_to_packet(&packet, sizeof(packet), peer);
96 		if (wg_noise_handshake_begin_session(&peer->handshake,
97 						     &peer->keypairs)) {
98 			wg_timers_session_derived(peer);
99 			wg_timers_any_authenticated_packet_traversal(peer);
100 			wg_timers_any_authenticated_packet_sent(peer);
101 			atomic64_set(&peer->last_sent_handshake,
102 				     ktime_get_coarse_boottime_ns());
103 			wg_socket_send_buffer_to_peer(peer, &packet,
104 						      sizeof(packet),
105 						      HANDSHAKE_DSCP);
106 		}
107 	}
108 }
109 
110 void wg_packet_send_handshake_cookie(struct wg_device *wg,
111 				     struct sk_buff *initiating_skb,
112 				     __le32 sender_index)
113 {
114 	struct message_handshake_cookie packet;
115 
116 	net_dbg_skb_ratelimited("%s: Sending cookie response for denied handshake message for %pISpfsc\n",
117 				wg->dev->name, initiating_skb);
118 	wg_cookie_message_create(&packet, initiating_skb, sender_index,
119 				 &wg->cookie_checker);
120 	wg_socket_send_buffer_as_reply_to_skb(wg, initiating_skb, &packet,
121 					      sizeof(packet));
122 }
123 
124 static void keep_key_fresh(struct wg_peer *peer)
125 {
126 	struct noise_keypair *keypair;
127 	bool send = false;
128 
129 	rcu_read_lock_bh();
130 	keypair = rcu_dereference_bh(peer->keypairs.current_keypair);
131 	if (likely(keypair && READ_ONCE(keypair->sending.is_valid)) &&
132 	    (unlikely(atomic64_read(&keypair->sending.counter.counter) >
133 		      REKEY_AFTER_MESSAGES) ||
134 	     (keypair->i_am_the_initiator &&
135 	      unlikely(wg_birthdate_has_expired(keypair->sending.birthdate,
136 						REKEY_AFTER_TIME)))))
137 		send = true;
138 	rcu_read_unlock_bh();
139 
140 	if (send)
141 		wg_packet_send_queued_handshake_initiation(peer, false);
142 }
143 
144 static unsigned int calculate_skb_padding(struct sk_buff *skb)
145 {
146 	/* We do this modulo business with the MTU, just in case the networking
147 	 * layer gives us a packet that's bigger than the MTU. In that case, we
148 	 * wouldn't want the final subtraction to overflow in the case of the
149 	 * padded_size being clamped.
150 	 */
151 	unsigned int last_unit = skb->len % PACKET_CB(skb)->mtu;
152 	unsigned int padded_size = ALIGN(last_unit, MESSAGE_PADDING_MULTIPLE);
153 
154 	if (padded_size > PACKET_CB(skb)->mtu)
155 		padded_size = PACKET_CB(skb)->mtu;
156 	return padded_size - last_unit;
157 }
158 
159 static bool encrypt_packet(struct sk_buff *skb, struct noise_keypair *keypair)
160 {
161 	unsigned int padding_len, plaintext_len, trailer_len;
162 	struct scatterlist sg[MAX_SKB_FRAGS + 8];
163 	struct message_data *header;
164 	struct sk_buff *trailer;
165 	int num_frags;
166 
167 	/* Calculate lengths. */
168 	padding_len = calculate_skb_padding(skb);
169 	trailer_len = padding_len + noise_encrypted_len(0);
170 	plaintext_len = skb->len + padding_len;
171 
172 	/* Expand data section to have room for padding and auth tag. */
173 	num_frags = skb_cow_data(skb, trailer_len, &trailer);
174 	if (unlikely(num_frags < 0 || num_frags > ARRAY_SIZE(sg)))
175 		return false;
176 
177 	/* Set the padding to zeros, and make sure it and the auth tag are part
178 	 * of the skb.
179 	 */
180 	memset(skb_tail_pointer(trailer), 0, padding_len);
181 
182 	/* Expand head section to have room for our header and the network
183 	 * stack's headers.
184 	 */
185 	if (unlikely(skb_cow_head(skb, DATA_PACKET_HEAD_ROOM) < 0))
186 		return false;
187 
188 	/* Finalize checksum calculation for the inner packet, if required. */
189 	if (unlikely(skb->ip_summed == CHECKSUM_PARTIAL &&
190 		     skb_checksum_help(skb)))
191 		return false;
192 
193 	/* Only after checksumming can we safely add on the padding at the end
194 	 * and the header.
195 	 */
196 	skb_set_inner_network_header(skb, 0);
197 	header = (struct message_data *)skb_push(skb, sizeof(*header));
198 	header->header.type = cpu_to_le32(MESSAGE_DATA);
199 	header->key_idx = keypair->remote_index;
200 	header->counter = cpu_to_le64(PACKET_CB(skb)->nonce);
201 	pskb_put(skb, trailer, trailer_len);
202 
203 	/* Now we can encrypt the scattergather segments */
204 	sg_init_table(sg, num_frags);
205 	if (skb_to_sgvec(skb, sg, sizeof(struct message_data),
206 			 noise_encrypted_len(plaintext_len)) <= 0)
207 		return false;
208 	return chacha20poly1305_encrypt_sg_inplace(sg, plaintext_len, NULL, 0,
209 						   PACKET_CB(skb)->nonce,
210 						   keypair->sending.key);
211 }
212 
213 void wg_packet_send_keepalive(struct wg_peer *peer)
214 {
215 	struct sk_buff *skb;
216 
217 	if (skb_queue_empty(&peer->staged_packet_queue)) {
218 		skb = alloc_skb(DATA_PACKET_HEAD_ROOM + MESSAGE_MINIMUM_LENGTH,
219 				GFP_ATOMIC);
220 		if (unlikely(!skb))
221 			return;
222 		skb_reserve(skb, DATA_PACKET_HEAD_ROOM);
223 		skb->dev = peer->device->dev;
224 		PACKET_CB(skb)->mtu = skb->dev->mtu;
225 		skb_queue_tail(&peer->staged_packet_queue, skb);
226 		net_dbg_ratelimited("%s: Sending keepalive packet to peer %llu (%pISpfsc)\n",
227 				    peer->device->dev->name, peer->internal_id,
228 				    &peer->endpoint.addr);
229 	}
230 
231 	wg_packet_send_staged_packets(peer);
232 }
233 
234 static void wg_packet_create_data_done(struct sk_buff *first,
235 				       struct wg_peer *peer)
236 {
237 	struct sk_buff *skb, *next;
238 	bool is_keepalive, data_sent = false;
239 
240 	wg_timers_any_authenticated_packet_traversal(peer);
241 	wg_timers_any_authenticated_packet_sent(peer);
242 	skb_list_walk_safe(first, skb, next) {
243 		is_keepalive = skb->len == message_data_len(0);
244 		if (likely(!wg_socket_send_skb_to_peer(peer, skb,
245 				PACKET_CB(skb)->ds) && !is_keepalive))
246 			data_sent = true;
247 	}
248 
249 	if (likely(data_sent))
250 		wg_timers_data_sent(peer);
251 
252 	keep_key_fresh(peer);
253 }
254 
255 void wg_packet_tx_worker(struct work_struct *work)
256 {
257 	struct crypt_queue *queue = container_of(work, struct crypt_queue,
258 						 work);
259 	struct noise_keypair *keypair;
260 	enum packet_state state;
261 	struct sk_buff *first;
262 	struct wg_peer *peer;
263 
264 	while ((first = __ptr_ring_peek(&queue->ring)) != NULL &&
265 	       (state = atomic_read_acquire(&PACKET_CB(first)->state)) !=
266 		       PACKET_STATE_UNCRYPTED) {
267 		__ptr_ring_discard_one(&queue->ring);
268 		peer = PACKET_PEER(first);
269 		keypair = PACKET_CB(first)->keypair;
270 
271 		if (likely(state == PACKET_STATE_CRYPTED))
272 			wg_packet_create_data_done(first, peer);
273 		else
274 			kfree_skb_list(first);
275 
276 		wg_noise_keypair_put(keypair, false);
277 		wg_peer_put(peer);
278 	}
279 }
280 
281 void wg_packet_encrypt_worker(struct work_struct *work)
282 {
283 	struct crypt_queue *queue = container_of(work, struct multicore_worker,
284 						 work)->ptr;
285 	struct sk_buff *first, *skb, *next;
286 
287 	while ((first = ptr_ring_consume_bh(&queue->ring)) != NULL) {
288 		enum packet_state state = PACKET_STATE_CRYPTED;
289 
290 		skb_list_walk_safe(first, skb, next) {
291 			if (likely(encrypt_packet(skb,
292 					PACKET_CB(first)->keypair))) {
293 				wg_reset_packet(skb);
294 			} else {
295 				state = PACKET_STATE_DEAD;
296 				break;
297 			}
298 		}
299 		wg_queue_enqueue_per_peer(&PACKET_PEER(first)->tx_queue, first,
300 					  state);
301 
302 	}
303 }
304 
305 static void wg_packet_create_data(struct sk_buff *first)
306 {
307 	struct wg_peer *peer = PACKET_PEER(first);
308 	struct wg_device *wg = peer->device;
309 	int ret = -EINVAL;
310 
311 	rcu_read_lock_bh();
312 	if (unlikely(READ_ONCE(peer->is_dead)))
313 		goto err;
314 
315 	ret = wg_queue_enqueue_per_device_and_peer(&wg->encrypt_queue,
316 						   &peer->tx_queue, first,
317 						   wg->packet_crypt_wq,
318 						   &wg->encrypt_queue.last_cpu);
319 	if (unlikely(ret == -EPIPE))
320 		wg_queue_enqueue_per_peer(&peer->tx_queue, first,
321 					  PACKET_STATE_DEAD);
322 err:
323 	rcu_read_unlock_bh();
324 	if (likely(!ret || ret == -EPIPE))
325 		return;
326 	wg_noise_keypair_put(PACKET_CB(first)->keypair, false);
327 	wg_peer_put(peer);
328 	kfree_skb_list(first);
329 }
330 
331 void wg_packet_purge_staged_packets(struct wg_peer *peer)
332 {
333 	spin_lock_bh(&peer->staged_packet_queue.lock);
334 	peer->device->dev->stats.tx_dropped += peer->staged_packet_queue.qlen;
335 	__skb_queue_purge(&peer->staged_packet_queue);
336 	spin_unlock_bh(&peer->staged_packet_queue.lock);
337 }
338 
339 void wg_packet_send_staged_packets(struct wg_peer *peer)
340 {
341 	struct noise_symmetric_key *key;
342 	struct noise_keypair *keypair;
343 	struct sk_buff_head packets;
344 	struct sk_buff *skb;
345 
346 	/* Steal the current queue into our local one. */
347 	__skb_queue_head_init(&packets);
348 	spin_lock_bh(&peer->staged_packet_queue.lock);
349 	skb_queue_splice_init(&peer->staged_packet_queue, &packets);
350 	spin_unlock_bh(&peer->staged_packet_queue.lock);
351 	if (unlikely(skb_queue_empty(&packets)))
352 		return;
353 
354 	/* First we make sure we have a valid reference to a valid key. */
355 	rcu_read_lock_bh();
356 	keypair = wg_noise_keypair_get(
357 		rcu_dereference_bh(peer->keypairs.current_keypair));
358 	rcu_read_unlock_bh();
359 	if (unlikely(!keypair))
360 		goto out_nokey;
361 	key = &keypair->sending;
362 	if (unlikely(!READ_ONCE(key->is_valid)))
363 		goto out_nokey;
364 	if (unlikely(wg_birthdate_has_expired(key->birthdate,
365 					      REJECT_AFTER_TIME)))
366 		goto out_invalid;
367 
368 	/* After we know we have a somewhat valid key, we now try to assign
369 	 * nonces to all of the packets in the queue. If we can't assign nonces
370 	 * for all of them, we just consider it a failure and wait for the next
371 	 * handshake.
372 	 */
373 	skb_queue_walk(&packets, skb) {
374 		/* 0 for no outer TOS: no leak. TODO: at some later point, we
375 		 * might consider using flowi->tos as outer instead.
376 		 */
377 		PACKET_CB(skb)->ds = ip_tunnel_ecn_encap(0, ip_hdr(skb), skb);
378 		PACKET_CB(skb)->nonce =
379 				atomic64_inc_return(&key->counter.counter) - 1;
380 		if (unlikely(PACKET_CB(skb)->nonce >= REJECT_AFTER_MESSAGES))
381 			goto out_invalid;
382 	}
383 
384 	packets.prev->next = NULL;
385 	wg_peer_get(keypair->entry.peer);
386 	PACKET_CB(packets.next)->keypair = keypair;
387 	wg_packet_create_data(packets.next);
388 	return;
389 
390 out_invalid:
391 	WRITE_ONCE(key->is_valid, false);
392 out_nokey:
393 	wg_noise_keypair_put(keypair, false);
394 
395 	/* We orphan the packets if we're waiting on a handshake, so that they
396 	 * don't block a socket's pool.
397 	 */
398 	skb_queue_walk(&packets, skb)
399 		skb_orphan(skb);
400 	/* Then we put them back on the top of the queue. We're not too
401 	 * concerned about accidentally getting things a little out of order if
402 	 * packets are being added really fast, because this queue is for before
403 	 * packets can even be sent and it's small anyway.
404 	 */
405 	spin_lock_bh(&peer->staged_packet_queue.lock);
406 	skb_queue_splice(&packets, &peer->staged_packet_queue);
407 	spin_unlock_bh(&peer->staged_packet_queue.lock);
408 
409 	/* If we're exiting because there's something wrong with the key, it
410 	 * means we should initiate a new handshake.
411 	 */
412 	wg_packet_send_queued_handshake_initiation(peer, false);
413 }
414